Resource-Efficient Time-Series Forecasting of Displacement Imagery using Koopman Autoencoders

Takayuki Shinohara, Hidetaka Saomoto

Keywords: Koopman operator, Time series InSAR, Ground Displacement, AutoEncoder

Abstract: Accurate yet lightweight forecasting of ground displacement is vital for real-time hazard response. We propose the Koopman Operator Autoencoder (KOA), a deep model that embeds a linear, physics-inspired Koopman operator into its latent space. A compact CNN encoder compresses each SBAS-InSAR frame; temporal evolution is then propagated by a single linear map, slashing parameter count and FLOPs relative to Transformer-style networks. Trained on nationwide Japanese SBAS archives and evaluated on unseen regions (Turkey, Italy, Hawaii), KOA matches state-of-the-art accuracy while cutting computational cost by orders of magnitude. This efficiency makes KOA practical for deployment on modest hardware in operational monitoring systems.

Submission Number: 2